Evaluating the quality of Machine Translation (MT) for low-resource languages is a challenging cross-lingual problem in Natural Language Processing. Quality Estimation (QE) aims to predict MT quality in the absence of reference translations. This thesis investigates QE through multilingual modelling, architectural adaptations, novel data curation and efficient learning strategies across nine low-resource language pairs. First, this thesis analyses how multilingual training influences QE performance, showing that combining linguistically related languages yields consistent improvements over training on a single language pair. The research then examines the suitability of large language models (LLMs) for QE in zero-shot, in-context learning, and fine-tuned settings. While LLMs show broad generalisation ability, empirical analysis reveals limitations in low-resource languages, including inconsistencies in tokenisation and weaker cross-lingual alignment. Results indicate that prompt-based adaptation alone is insufficient for reliable QE for low-resource languages and motivate architectural refinement. Building on this finding, the thesis proposes ALOPE, a regression-aligned framework for adapting decoder-based LLMs by leveraging informative intermediate Transformer layers for QE. Combined with lightweight regression heads and parameter-efficient techniques, ALOPE enables smaller open-weight LLMs to achieve state-of-the-art performance on multiple language pairs while operating within memory and compute constraints. Further, this research extends QE beyond scalar prediction by introducing a dataset enriched with human-annotated error descriptions that provide contextual supervision. Building on this, we propose ALOPE-RL, a policy-based reinforcement learning framework that optimises quantised LLMs using multi-component rewards, achieving state-of-the-art performance and remaining effective under extreme data constraints. This thesis shows that QE for low-resource languages can be performed through linguistically grounded multilingual transfer, regression-aligned adaptation, and error-aware reinforcement learning, for practical translation evaluation. All datasets, models, and code are publicly released.
Archchana Sindhujan· University of Surrey Open Re...· 0 citations
An independent, clean-room reimplementation of a self-model loop architecture for language-model agents (the "AC1 loop" described in Lark Laflamme's 2026 AC1-LLM / Laflamme-3T essays). The architecture wraps an LLM with a Bayesian belief over the agent's own interaction stance, re-injected into generation each turn alongside a hidden deliberative monologue. The originating write-ups report strong effects but omit the controls needed to separate the contribution of the state's content from the mere presence of an instruction, an honest task baseline, or the intrinsic inertia of the estimator. This work supplies those controls across six experiments: a structured-placebo ablation (E1), a clamped partial factorial over state x gate x monologue x strategy (E2), hidden-stance inference against an honest single-shot LLM baseline (E3), and closed-loop dynamics driven against an OpenAI-compatible chat endpoint with decay-free and state-decoupled null controls (E4–E6). In small synthetic experiments on one or a few model endpoints, mode-labelled prompt additions reliably changed output length and question use, most strongly when the posterior was confident and with the private monologue as the leading but not isolated factor. A classifier-plus-accumulator inferred synthetic modes well above chance without an accuracy advantage over an honest single-shot baseline. Under a cyclic numerical drive the closed loop produced substantial finite-horizon path dependence, most of it accumulator arithmetic, with the size and mechanism of any additional feedback contribution left uncertain; a 100-turn basin test found no evidence of bistability, with relaxation still in progress. Every tested claim reproduces in a "true-but-softer" form once the missing controls are added. Scope is strictly the loop's measurable behaviour; no claim is made about consciousness or any Psi-threshold. Code and raw result data are released (MIT). Version 1.1 revises v1.0 after an external validity review: corrections of fact (E2 is 30 configurations / 180 replies and a partial factorial; the "numbers-only" control is renamed strategy-stripped; the gate channel is substantial, not minor; "loop area" is a mean vertical separation), an E1 reproducibility caveat, and claims softened to what the designs support. See the paper's Revision history. No data were re-run.
Grant Wilson· Zenodo (CERN European Organi...· 0 citations
Fil-C showed that C can be made memory-safe at run time by giving every pointer acapability and checking it on every access. We take its diagnostic ambitions and put themsomewhere else: in a compiler small enough for a person, or a language model, to hold inview at once, and with a different goal. Crust’s --mem-safe is a test-time tool rather thana deployment mitigation. The identical source builds with the flag for the test suite andwithout it for release, and the release binary carries no runtime at all.The difference in setting changes what is possible. Because the check is inserted by thecompiler on its own intermediate language, the compiler can also decide not to insert it.Three proof rules—redundancy by local value numbering, constant-offset bounds againsta statically known allocation, and loop-carried index ranges from a dominating guard—together with hoisting of the remaining shadow bookkeeping out of the loop, took a 2M-iteration array loop from 31× overhead under uniform instrumentation to 1.05×, with noruntime call left in the loop body. Fil-C cannot do this: its check is a per-pointer test withno whole-program view to prove anything away.Because the bounds live in a side table keyed by address rather than in the pointer,instrumented and uninstrumented translation units link freely. That is what makes a secondtier possible: --mem-safe=cpp checks only code lowered from the C++ subset and leavesalready-audited C at full speed, distinguishing the two by the file name on each IL command’ssource range.We then argue the larger point: that a memory-safety mechanism belongs in a toolchainthat can be modified and re-verified in the time a test run takes, and that a fork like Fil-C126,031 commits behind its upstream is not such a toolchain.
Brent Hartshorn· Zenodo (CERN European Organi...· 0 citations
The rapid advancement of information technology has encouraged Micro, Small, and Medium Enterprises (MSMEs) to adopt digital solutions to improve business performance and customer service. Stasiun Jajanan is a culinary MSME that offers various fast-food products, including kebabs, burgers, fried tapioca fritters, Indonesian fried tofu meatballs, and beverages. Promotional activities, order processing, and transaction recording are still conducted manually, resulting in limited promotional reach, inefficient data management, and potential recording errors. This study aims to design and develop a web-based restaurant information system to support business promotion, online ordering, and transaction management for Stasiun Jajanan. The system was developed using the System Development Life Cycle (SDLC) with the Waterfall model, consisting of requirement analysis, system design, development, and testing. Unified Modeling Language (UML) was utilized to model the system through use case, activity, sequence, and class diagrams. The developed system provides several features, including menu management, online ordering, payment processing, transaction management, business information services, and WhatsApp integration for customer communication. System testing was conducted using the Black Box Testing method to verify that all functionalities operated according to the specified requirements. The testing results indicate that the developed system successfully supports menu management, customer ordering, payment processing, and transaction recording. The system improves service efficiency, reduces manual recording errors, and provides customers with easier access to product information and ordering services. The proposed system can therefore support the digital transformation and business growth of culinary MSMEs.
Agus Ganda Permana, Dewi Safitriani, Devit Setiono· TEPIAN· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Pediatric infectious diseases are moving into a phase where the old boundaries of the specialty just don't hold up anymore. A febrile child might be the first visible point in a much longer chain.In 2024, a dengue outbreak in Fano, Italy, hit 199 locally acquired cases, and France logged 83 autochthonous infections-numbers that would have seemed out of place in continental Europe a decade ago [1]. For pediatric infectious disease specialists, these aren't just epidemiological oddities; they show how climate, competent vectors, human mobility, and delayed recognition can all converge fast. But the clinical encounter only helps with earlier outbreak detection, safer prescribing, and prevention if clinical data actually link up with microbiological, entomological, environmental, and public health signals. So the real challenge isn't just responding to more pathogens-it's building a child-centered learning system that connects human, animal, and environmental health through responsible digital transformation.One Health gives us the conceptual framework for this shift. It starts from the idea that the health of people, animals, plants, and ecosystems is all interdependent, and the Quadripartite One Health Joint Plan of Action flags zoonoses, vector-borne diseases, food safety, antimicrobial resistance (AMR), and environmental health as connected priorities [2,3]. Yet One Health often stays an aspiration rather than a routine part of pediatric practice. Digital transformation could supply the connective tissue: interoperable surveillance, geospatial analysis, pathogen genomics, environmental sensors, electronic health records, and artificial intelligence (AI) can turn fragmented observations into shared situational awareness. Technology, though, isn't the end goal. Its value lies in enabling earlier, more equitable, and more sustainable decisions for children.Children aren't just small adults, either biologically or in terms of exposure. Immune maturation, age-dependent clinical phenotypes, vaccine schedules, developmental behavior, and weightbased treatment all shape susceptibility and outcomes. Their environments are distinct too: households, schools, childcare facilities, playgrounds, farms, and urban transport determine contact with pathogens, animals, pollutants, and vectors. Infants and young children have higher ventilation relative to body size, frequent hand-to-mouth activity, and little control over their surroundings. Social deprivation, displacement, and unequal access to vaccination, sanitation, diagnostics, and digital infrastructure further concentrate risk.Climate change makes these interdependencies more visible by the year. Temperature, rainfall, drought, flooding, and ecosystem disruption influence the geography and seasonality of vectorborne, waterborne, foodborne, and respiratory infections. Europe's experience with locally acquired arboviruses shows how competent vectors, favorable weather, and imported index cases can converge to create autochthonous transmission, while pediatric infections may stay underdiagnosed because they're mild or nonspecific [1]. Air pollution adds another layer: it can damage epithelial barriers, modify immune responses, and raise the risk or severity of respiratory infections in children [4]. Environmental exposures may also reshape the respiratory microbiota and favor colonization by potential pathogens [5]. The gut and airway microbiomes, then, should be seen as dynamic interfaces between the child, antimicrobials, nutrition, pathogens, and the environment-not isolated laboratory curiosities.Current surveillance systems are usually separated by institution, discipline, and administrative level. Pediatric clinical data sit in hospital records; microbiological results in laboratory systems; antimicrobial consumption in pharmacy databases; vaccination in registries; animal infections in veterinary networks; and climate, pollution, wastewater, and vector data in other agencies. A digitally enabled One Health model wouldn't indiscriminately centralize everything. It would set up interoperable, privacy-preserving pathways that allow relevant signals to be combined across space and time.Such infrastructure could detect an unusual pediatric syndrome, connect it with a veterinary or environmental alert, map it against weather and mobility patterns, and quickly return guidance to clinicians. Wastewater and environmental sequencing could complement clinical testing, while genomic epidemiology could distinguish community transmission from healthcare-associated clusters. Federated analysis could let institutions collaborate without transferring identifiable child-level data. Dashboards could provide age-stratified incidence and resistance patterns, and automated alerts could support-rather than replace-public health judgment. This direction aligns with the WHO global strategy on digital health, but pediatric implementation needs explicit age-sensitive priorities [6].AI may enhance forecasting, image and signal interpretation, triage, diagnostic support, and antimicrobial decision-making. Still, the gap between model development and clinical benefit remains wide. A systematic review of predictive AI implemented in pediatric practice found few real-world implementations and inconsistent evaluation of clinical, workflow, and human outcomes [7]. In pediatric infectious diseases, digital technologies show promise for monitoring, access, communication, and timely intervention, but infrastructure and implementation gaps persist [8]. Future research has to move beyond retrospective accuracy. Models should be externally validated across ages, settings, ethnic and socioeconomic groups; compared with existing care; evaluated prospectively; monitored for performance drift; and designed with clinicians, families, and public health professionals.Digital transformation can be most immediately useful where uncertainty drives unnecessary treatment. Rapid molecular tests, multiplex platforms, biomarkers, and metagenomic nextgeneration sequencing can shorten diagnostic pathways and reveal uncommon or mixed infections. Recent pediatric evidence suggests that metagenomic sequencing offers high sensitivity and frequently changes antimicrobial management, although cost, interpretation, standardization, and contamination remain challenges [8]. Diagnostic innovation therefore has to be coupled to diagnostic stewardship: selecting the right test, for the right child, at the right time, and interpreting it in its clinical and epidemiological context.Antimicrobial stewardship should likewise evolve from periodic prescription review to a continuously learning One Health function. Clinical decision support could integrate age, syndrome, allergy history, previous microbiology, local resistance, drug availability, pharmacokinetics, and organ function. At population level, pediatric prescribing could be analyzed alongside resistance and antimicrobial use in animals and, when feasible, environmental reservoirs. European integrated analyses already demonstrate the importance of coordinated surveillance across human and food-producing animal sectors and call for inclusion of environmental data [9]. The next step is making this information actionable at the bedside without increasing alert fatigue or inequity.The same approach applies to healthcare-associated infections. Neonatal and pediatric intensive care, oncology, transplantation, surgery, and long-term technology dependence create distinctive risks. Electronic surveillance can combine device days, microbiology, antibiotic exposure, staffing, movement through wards, and environmental sampling to identify preventable harm earlier. Genomic data can clarify transmission, while real-time feedback can improve adherence to infection-prevention bundles. These tools must be judged by reduced infections, antimicrobial use, length of stay, and family burden-not merely by technical detection rates.One Health and digital transformation also open up a chance to study mechanisms rather than correlations alone. Longitudinal pediatric cohorts can combine clinical phenotypes with gut and respiratory microbiome profiles, pathogen detection, antimicrobial exposure, diet, vaccination, air quality, temperature, housing, green space, and social determinants. Wearable and geospatial technologies may improve exposure measurement, while multi-omics and causal inference can help identify pathways connecting pollution or climate-related exposures to mucosal immunity and infection. This agenda requires restraint as well as ambition. Microbiome signatures can be population-and context-specific; sequencing detects nucleic acid rather than necessarily viable disease-causing organisms; environmental measurements are prone to misclassification; and highly dimensional models can generate convincing but non-reproducible associations. Studies should therefore prioritize standardized sampling, transparent analytic plans, clinically meaningful endpoints, diverse populations, and independent validation. Interventions-such as reducing harmful environmental exposure, improving ventilation, optimizing antibiotic use, or restoring microbial ecosystems-should ultimately be tested for their ability to improve child health.Emerging and re-emerging infections are a defining challenge for pediatric infectious diseases.Zoonotic spillover, vector expansion, climate disruption, conflict, displacement, declining vaccine coverage, and fragile health systems can transform a local signal into a regional or global emergency. Children may be disproportionately affected yet remain under-represented in surveillance, therapeutic trials, and emergency preparedness. Readiness therefore requires pediatric protocols, pre-approved adaptive trial platforms, biobanking and genomic capacity, rapid data sharing, and pathways that can pivot from endemic diseases to outbreaks without abandoning routine care.Preparedness can't be separated from access. Infectious diseases remain leading causes of childhood death, while many low-and middle-income countries (LMICs) face limited access to laboratories, imaging, oxygen, vaccines, child-appropriate formulations, and effective antimicrobials [10]. Innovations developed during emergencies have too often reached the populations bearing the greatest burden late, at unaffordable prices, or not at all. Equity must be designed across the full pathway-from priority setting and product development to procurement, delivery, and post-implementation evaluation. Digital tools should function with intermittent connectivity, limited computing resources, and local languages rather than widening the digital divide.Research in LMICs must also move beyond extractive models. Institutions and investigators in affected countries should lead agenda setting, governance, analysis, authorship, and translation into policy; communities should participate from protocol design onward; and data and biological samples should be governed through fair, transparent agreements. Sustainable investment in laboratories, ethics and regulatory systems, trial networks, data stewardship, and research careers is itself a preparedness intervention. WHO guidance calls for country-led, continuously functional clinical-trial ecosystems and highlights the persistent underrepresentation of children [11]. Pediatric innovation should be judged not only by discovery, but by whether affordable diagnostics, vaccines, medicines, and evidence reach every child who needs them.The greatest risk is that digital One Health becomes a technically sophisticated system that doesn't serve the children at highest risk. Data gaps aren't random: populations with limited healthcare access, unstable housing, migration, conflict, or weak connectivity are often least visible to digital systems. Algorithms trained on data from tertiary hospitals or high-income settings may underperform elsewhere. Surveillance may also create harms through stigmatization of communities, locations, occupations, or animal contacts.Children merit heightened safeguards because they can't always provide consent, their data may remain identifiable across a lifetime, and decisions made today can shape future opportunities. WHO guidance emphasizes that AI for health must be governed by ethics, human rights, transparency, accountability, and protection of autonomy [12]. Pediatric systems should apply data minimization, clear purpose limitation, age-appropriate assent, meaningful caregiver engagement, cybersecurity, auditability, and defined human oversight. Families and young people should participate in determining which uses of data are acceptable and valuable.Equity must be an outcome, not a generic principle. Every digital intervention should report who was included, who was excluded, who benefited, and who carried additional burdens. Open standards, adaptable low-resource tools, multilingual communication, and capacity building are essential. Investment is also needed in a workforce fluent across pediatrics, infectious diseases, microbiology, epidemiology, veterinary and environmental sciences, data science, behavioral science, ethics, and implementation research. No single profession can deliver this transformation.The pediatric infectious diseases community should make five commitments. First, we should build interoperable, age-stratified One Health surveillance that connects clinical, laboratory, genomic, antimicrobial, veterinary, vector, wastewater, climate, and pollution signals. Second, we should evaluate digital tools by prospective clinical and public health impact, usability, sustainability, and equity-not by algorithmic performance alone. Third, we should integrate diagnostic stewardship, antimicrobial stewardship, and infection prevention into the same learning system. Fourth, we should establish longitudinal research platforms linking the microbiome and exposome to infection while moving toward testable preventive interventions.Fifth, we should create a pediatric digital governance framework in which children's rights, family participation, transparency, and accountability are designed in from the start.These commitments also define a broad research scope for the specialty: emerging and vectorborne infections; zoonoses and foodborne disease; AMR; healthcare-associated infections; vaccines; microbiome science; environmental determinants; digital diagnostics; AI; implementation; and health equity. The unifying question is whether research can connect these domains in ways that improve decisions for children.The future of pediatric infectious diseases will be determined not only by which pathogens emerge, but by whether we can recognize and respond to connected threats before they become crises. One Health supplies the necessary systems perspective; digital transformation provides tools to make that perspective operational. Neither is sufficient alone. One Health without interoperable data risks remaining rhetorical, while digital innovation without ecological context, clinical validation, and ethical governance risks producing faster fragmentation.Our grand challenge is to create a trustworthy, equitable learning ecosystem in which information flows responsibly from the child's bedside to the laboratory, community, environment, and public health system-and back again as timely action. Pediatric infectious disease specialists should help lead this transition because we understand both the urgency of infection and the long horizon of childhood. Protecting children today also protects the microbial, social, and environmental conditions in which future generations will live.
Daniele Zama· Frontiers in Pediatrics· 0 citations
Financial disclosures contain rich narrative information about firm performance, but their length, specialized terminology, and target-dependent language make sentence-level analysis challenging. Existing financial sentiment methods typically assess overall positive or negative tone, rather than stance toward specific financial targets. In this study, we introduce a sentence-level benchmark for target-specific financial stance detection in Form 10-K Management’s Discussion and Analysis (MD&A) sections and quarterly earnings call transcripts (ECTs). The benchmark focuses on three financial targets–debt, earnings per share (EPS), and sales–and assigns each target-relevant sentence a positive, negative, or neutral stance label. The corpus is drawn from five public companies. The study is intended as a benchmark for NLP methods and does not support broad claims about financial disclosures in general due to the limited set of companies and financial targets. For scalability, the training split is labeled using ChatGPT-o3-pro, while the held-out test split is independently annotated by human annotators and adjudicated to form a human-consensus gold standard. Using this benchmark, we evaluate four contemporary large language models under zero-shot, few-shot, Chain-of-Thought, and document-context prompting conditions. Model outputs are assessed using accuracy, macro-averaged and weighted precision, recall, and F1, with paired statistical tests used to evaluate performance differences. Results show that LLMs can provide useful baselines for low-label, target-specific financial stance detection, but performance varies across models, document types, targets, and prompting strategies. GPT-4.1-mini and Gemma4-31B generally achieved the strongest overall performance, followed by Llama 3.3-70B and Mistral Small 3.2-24B, although no single model or prompting strategy dominated across all settings. The no-document-context condition achieved the highest Macro-F1 score. However, this result is best interpreted as alignment with the sentence-level annotation protocol, in which annotators did not have access to the full document context, rather than as evidence that document context is generally unhelpful. These findings highlight both the promise and the remaining limitations of LLM-based financial stance detection, particularly the need for broader industry coverage and economic validation.
Nikesh Gyawali, Doina Caragea, Alex Vasenkov et al.· Scientific Reports· 0 citations
Abstract Background To descriptively evaluate short-term (48–72 h postoperatively) neurological functional outcomes in patients undergoing awake craniotomy for intracranial lesions located within eloquent brain areas, and conduct univariate exploratory analyses of clinical factors correlated with acute postoperative neurological deterioration. Secondary exploratory survival analyses are included solely for hypothesis generation; all survival-related findings carry no definitive clinical prognostic interpretation given the study’s critical limitations of small sample size, pathological heterogeneity, absent long-term functional follow-up, and missing key perioperative covariates. Methods A single-center descriptive retrospective analysis was performed on 30 patients who received awake craniotomy between January 2019 and November 2024. The primary predefined endpoint was new or exacerbated neurological deficits within 48–72 h after surgery, quantified via Karnofsky Performance Status (KPS), Glasgow Coma Scale (GCS), Mini-Mental State Examination (MMSE), and manual muscle strength testing. Early postoperative dysfunction was categorized into three distinct phenotypes: transient global systemic functional decline, focal motor neurological deficits, and isolated language impairment (aphasia). Secondary exploratory endpoints included overall survival and all-cause mortality, with follow-up completed by February 28, 2025. Univariate logistic regression screened correlates of acute neurological decline; Kaplan–Meier survival curves, log-rank subgroup comparisons, and Cox proportional hazards regression were applied for exploratory survival modeling. All survival analyses are interpreted as purely descriptive and hypothesis-generating, with no confirmatory clinical value; the specific constraints are addressed in the Limitations section. Results The cohort contained 30 patients, with an awake craniotomy intraoperative mapping completion rate of 90.0% (27/30). Within 48–72 h postoperatively, 30.0% (9/30) of patients developed new or aggravated acute neurological dysfunction. Most acute impairments reflected transient global functional decline, including KPS reduction (30.0%) and MMSE cognitive decline (26.7%), while permanent focal injury surrogates (new motor weakness 16.7%; new-onset aphasia 10.0%) were far less frequent. We emphasize that early KPS/MMSE reductions largely stem from temporary perioperative cerebral edema, postoperative fatigue, and residual sedation recovery rather than irreversible focal brain injury. Patients with successful intraoperative awake mapping exhibited a numerically lower rate of acute neurological decline (25.9%) versus patients with mapping failure (66.7%), though this trend did not reach statistical significance ( P = 0.20). Univariate logistic regression identified no statistically significant independent predictors of acute postoperative neurological deficits. Median overall survival for the entire cohort was 43.0 months; all subgroup survival comparisons and Cox regression hazard models yielded non-significant results, with no reliable prognostic factors identified in this limited cohort. Conclusions Awake craniotomy combined with direct electrical stimulation mapping is technically feasible and effectively limits permanent language dysfunction in patients with eloquent cortex lesions. Transient short-term global functional deterioration is common in the immediate postoperative window, driven by reversible perioperative physiological disturbances rather than definitive surgical neurological injury. Owing to the retrospective single-center design, small sample size, heterogeneous intracranial pathology, absence of long-term (≥ 3-month) functional outcome assessments, lack of a general anesthesia control group, and unmeasured critical covariates (including extent of resection), all survival and mortality-related analyses are strictly exploratory and cannot support clinical prognostic judgements. Large, pathology-stratified controlled prospective studies incorporating serial long-term neurological evaluations are required to validate definitive risk factors and long-term functional benefits of awake craniotomy.
Yajie Song, Kaiwen Yang, Yushun Zhang et al.· Chinese Neurosurgical Journa...· 0 citations
Digital assistive technology (AT) is now delivered through a variety of platforms, including smartphones, cloud services, dedicated devices, and embedded systems rather than a single dominant form factor. However, in low- and middle-income countries (LMICs), affordability, connectivity, electrical reliability, service capacity, and uneven language infrastructure continue to determine which digital pathways can be deployed and maintained at scale. This perspective argues that market-shaping strategies should prioritise the development of shared digital infrastructure beneath the application layer; this includes representative datasets and language resources, reusable models and reference toolchains, and deployment pathways tailored to resource-constrained environments. It proposes a three-part Market Stack comprising governed data and language resources, reusable core models and reference pipelines, and localised commercialisation and service delivery by small and medium enterprises. The central argument is not to replace smartphones but to ensure that digital AT markets in constrained settings work best when innovators and procurers can choose among smartphone, dedicated, embedded, and hybrid solutions. Therefore, funders and policymakers should view data, language resources, and foundational tooling as public-interest investments that widen the range of assistive products available in LMIC markets.
Will Wade, Sijung Hu, James Fleming· Frontiers in Rehabilitation...· 0 citations
Large Language Models have advanced natural language generation, but they often produce outputs that are grammatically correct yet factually incorrect or misleading. This issue, commonly known as hallucination, reduces the reliability of such systems, especially in domains such as law, medicine, journalism, and education, where factual accuracy is important. This work presents a systematic low-resource framework for hallucination detection across three natural language processing tasks, including machine translation, definition modeling, and paraphrase generation. The experiments are conducted on the SemEval 2024 Task 6 SHROOM dataset, which consists of a small labeled set and a larger set of unlabeled data across three text generation tasks. An iterative self-training approach based on transformer models is employed, where a small manually labeled set is expanded using confidence-based pseudo-labeling. Active learning, ensemble methods and a Query by Committee strategy are used to guide sample selection and improve stability under limited supervision. The models are evaluated using accuracy and Spearman correlation to capture classification performance and prediction consistency. The proposed setup achieves an accuracy of 0.78 and a Spearman correlation of 0.6626. In addition, QLoRA-based large language models are evaluated as lightweight baselines for comparison with the proposed framework. The work presents a lightweight, semi-supervised, and reproducible framework that achieves competitive performance without relying on resource-intensive models. It serves as a practical baseline, highlights key challenges, and provides a foundation for more advanced hallucination detection systems in real-world applications.
Objective To develop and internally validate a prototype multimodal artificial intelligence system for automated CEAP (Clinical, Etiological, Anatomical and Pathophysiological) classification of venous duplex ultrasound (VDUS) reports, integrating natural language processing of free-text components with computer vision analysis of hand-drawn anatomical diagrams. Methods Single centre retrospective observational study using routinely collected clinical data. One thousand consecutive venous duplex ultrasound reports from Cambridge University Hospitals NHS Foundation Trust, UK (July 2024 – May 2025) were labelled according to the CEAP classification, excluding the Etiological component, which could not be reliably determined from duplex reports alone. Transfer learning was applied using ClinicalBERT for text and MobileNetV3 for diagrammatic data. Clinical classes were predicted from request line text. Text- and image-based pathophysiological models were developed for four anatomical territories (Great Saphenous Vein, Small Saphenous Vein, Deep system, Perforators), combined using late fusion with probability averaging. Results The clinical CEAP model achieved accuracy of 0.91, macro-F1 of 0.82, and macro-AUC of 0.98. Pathophysiological prediction varied, with text models broadly outperforming image models. Fusion yielded heterogeneous benefits, improving SSV performance but reducing Deep system accuracy. The performance of the final pathophysiological CEAP fusion models varied across anatomical territories: accuracy ranged from 0.70–0.92 and macro-AUC from 0.80–0.92. Conclusion This study demonstrates the feasibility of automated CEAP classification from VDUS reports. Despite class imbalance affecting minority class predictions, the strong discriminatory performance validates this multimodal ML model for extracting clinically meaningful information from real-world data. This approach offers potential, pending external validation, to streamline vascular services through automated triage and guideline-compliant decision making.
Joseph Cutteridge, Henry Bergman, Will Jackson et al.· Journal of Vascular Surgery...· 0 citations
Fil-C showed that C can be made memory-safe at run time by giving every pointer acapability and checking it on every access. We take its diagnostic ambitions and put themsomewhere else: in a compiler small enough for a person, or a language model, to hold inview at once, and with a different goal. Crust’s --mem-safe is a test-time tool rather thana deployment mitigation. The identical source builds with the flag for the test suite andwithout it for release, and the release binary carries no runtime at all.The difference in setting changes what is possible. Because the check is inserted by thecompiler on its own intermediate language, the compiler can also decide not to insert it.Three proof rules—redundancy by local value numbering, constant-offset bounds againsta statically known allocation, and loop-carried index ranges from a dominating guard—together with hoisting of the remaining shadow bookkeeping out of the loop, took a 2M-iteration array loop from 31× overhead under uniform instrumentation to 1.05×, with noruntime call left in the loop body. Fil-C cannot do this: its check is a per-pointer test withno whole-program view to prove anything away.Because the bounds live in a side table keyed by address rather than in the pointer,instrumented and uninstrumented translation units link freely. That is what makes a secondtier possible: --mem-safe=cpp checks only code lowered from the C++ subset and leavesalready-audited C at full speed, distinguishing the two by the file name on each IL command’ssource range.We then argue the larger point: that a memory-safety mechanism belongs in a toolchainthat can be modified and re-verified in the time a test run takes, and that a fork like Fil-C126,031 commits behind its upstream is not such a toolchain.
Brent Hartshorn· Zenodo (CERN European Organi...· 0 citations
Background The incidence of aneurysmal subarachnoid haemorrhage declined between 1980 and 2010, which coincided with a decline in smoking and prevalence of hypertension. We aimed to investigate whether the decrease in subarachnoid haemorrhage incidence is paralleled by declines in unruptured intracranial aneurysm (UIA) prevalence. Methods For this systematic review and meta-analysis, we searched Embase, PubMed, and Web of Science for articles published in any language from Jan 1, 2011 to Dec 31, 2025, and reassessed 68 articles published before March 1, 2011 from a 2011 systematic review and meta-analysis. Articles were eligible for inclusion if they used a cross-sectional or case–control design and provided the crude number of participants and those with UIA. We only included studies reporting numbers of UIA separately from ruptured aneurysms and with ten or more patients. Summary data were independently extracted by JD with AZ or CB and conflicts were resolved by GJER. The primary outcome was proportion of participants with UIA. Relative to a hypothetical reference population (mean age 50 years, 50% women, and no comorbidities), age and/or sex-adjusted prevalence ratios (PRs) for regions, comorbidities, and risk ratios (RRs) for female sex, smoking, and hypertension were estimated using generalised linear mixed models. A time trend analysis was done by binomial meta regression using the mid-year of data acquisition. We assessed the certainty of evidence using GRADE. The study was registered with PROSPERO, number CRD420261296728. Findings Our search screened 4708 studies. 67 reassessed and 95 newly identified articles, reporting on 316 131 participants and 11 822 people with UIAs, were included in our meta-analysis. In the reference population, the estimated prevalence of UIAs was 3·9% (95% CI 3·0–5·1). The prevalence of UIAs in individuals with atherosclerosis was 5·5% (4·7–6·4; 2229 of 40970 participants) and the adjusted PR was 1·3 (95% CI 0·8–2·0) compared with the reference population. For positive family history of aneurysmal subarachnoid haemorrhage (aSAH) or UIA, the UIA prevalence was 7·9% (5·6–11·1; 412 of 4252 participants) and the adjusted PR was 2·4 (0·5–11·2). For connective-tissue disorder, the UIA prevalence was 10·3% (6·5–16·0; 94 of 879 participants) and the adjusted PR was 3·9 (2·0–7·6). For autosomal dominant polycystic kidney disease (ADPKD), the UIA prevalence was 12·8% (9·2–17·6; 293 of 1990 participants) and the adjusted PR was 4·4 (1·5–12·6). RRs were for current smoking 1·4 (1·2–1·6; 798 of 27911 participants), for having hypertension 1·6 (1·5–1·7, 4043 of 83053 participants), and for female sex 1·9 (1·8–2·0; 3415 of 65020 women and 2122 of 76130 men). In studies on healthy individuals with MR angiography or CT angiography as imaging modality, the prevalence in 2016–2022 was 6·6% (6·3–6·8; 2904 of 41191 participants). The adjusted PR was 1·8 (1·1–2·8) for 2016–2022 versus 2002–2015. Prevalence of UIAs of 5 mm or larger was 0·7% (0·6–0·8) in 2002–2015 and 1·4% (1·0–1·9) in 2016–2022. The UIA prevalence did not differ between countries. τ 2 showed significant heterogeneity between studies. The certainty of the evidence ranged from very low to moderate. Interpretation Prevalence of UIA is increasing, particularly over the past two decades. This increase is only in part explained by improved detection of small UIAs and an ageing population, and other factors—such as environmental—are likely involved. Alongside patients with ADPKD and a positive family history of aSAH, patients with connective-tissue disorders had a higher prevalence of UIA than the reference population. Our findings warrant further investigation into the potential benefit of personalised screening and management strategies in groups at high risk for having UIAs. Funding None.
Judith Dremel, Viktoria Rücker, Camilo Badel et al.· The Lancet Neurology· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
